Answers are grounded in this paper and cite their sources.
Purpose Large multimodal models (LMMs) capable of interpreting medical images are increasingly accessible to patients, yet their reliability for retinal optical coherence tomography (OCT) interpretation remains uncertain. This study evaluated the performance of two state-of-the-art LMMs in detecting OCT features and generating diagnostic interpretations. Methods In this cross-sectional study, 330 publicly available retinal OCT B-scans spanning 12 diagnostic categories, including normal scans and common retinal diseases, were analyzed. Two commercially available LMMs (GPT 5.2 Pro and Gemini 3 Pro) were evaluated using zero-shot inference. Models were assessed separately for detection of nine predefined OCT features and for free-text diagnostic interpretation. Feature detection performance was measured using per-feature metrics and macro-averaged F1 scores, while diagnostic accuracy was assessed against expert-adjudicated reference standards. Results Gemini 3 Pro demonstrated higher feature detection performance than GPT 5.2 Pro (macro-averaged F1 score 0.558 vs. 0.472) and higher diagnostic accuracy (58.8% vs. 32.1%). Both models performed better for visually salient abnormalities but showed limited sensitivity for subtle structural features. Normal OCT scans were frequently misclassified as pathological, particularly by GPT 5.2 Pro, which failed to correctly identify any normal images. Recurrent, model-specific diagnostic biases and hallucinated diagnoses were observed. Conclusions Although LMMs can extract clinically relevant information from OCT scans, their diagnostic reliability remains limited, with frequent overcalling of pathology and systematic biases. Given their increasing availability to patients, these limitations raise concerns about potential misinterpretation outside supervised clinical settings. Careful validation and clear usage boundaries are needed before such tools can be safely integrated into patient-facing workflows.
No takes yet. Share an insight, caveat, or question.
Bellanda et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: